Future FSD update will park perfectly at superchargers

[HaulingAss]

If FSD recognizes it is ultimately being judged against the rules-based at-fault accident rate of human drivers, isn't its logical behavior to do its best to avoid being found at-fault, but also to undetectably trigger at-fault accidents among human drivers where possible?

Um, no. That demonstrates a basic lack of understanding of how FSD is trained and the kinds of capabilities it has.

Sponsored

 

[GuyV]

Um, no. That demonstrates a basic lack of understanding of how FSD is trained and the kinds of capabilities it has.

Haha, that's the thinking of people who believe they understand and control AI but somehow can't make it do what they want it to.

They try to massage the training data but don't have anywhere near the capability that they have given the AI to discern patterns. They even use AI generated synthetic data. They really can't know for sure the lessons they are teaching it. The results are trial and error, and they don't know how weird and dangerous a seemingly good version might be until they let it loose in the real world.

You know, they've told us all that, but didn't quite state it that way.

 

[HaulingAss]

The results are trial and error, and they don't know how weird and dangerous a seemingly good version might be until they let it loose in the real world.

Nope, it's a basic misunderstanding. They have already released it in the real world and we can collect data on the accident rate of each version. When it has a slight regression, they revert back to versions that were safer.

One of the biggest misunderstandings about autonomous driving is that if we don't know what's going on "under the covers" then it's scary. That's false because we can see the result and results are all that matter. It's called data. Similarly, we have a very poor understanding of what's really going on inside humans when they drive, sometimes the fail to see a car, motorcycle or bicyclist coming right towards them as they pull out. But we just accept these human inconsistencies as normal.

 

[GuyV]

Nope, it's a basic misunderstanding. They have already released it in the real world and we can collect data on the accident rate of each version. When it has a slight regression, they revert back to versions that were safer.

One of the biggest misunderstandings about autonomous driving is that if we don't know what's going on "under the covers" then it's scary. That's false because we can see the result and results are all that matter. It's called data. Similarly, we have a very poor understanding of what's really going on inside humans when they drive, sometimes the fail to see a car, motorcycle or bicyclist coming right towards them as they pull out. But we just accept these human inconsistencies as normal.

That data collection is after the fact and no one knows what crazy thing it might do in some situation until it encounters that situation and does that thing.

You are correct in that it is quite similar to our poor understanding of what's going on inside humans, and as I said before, that's a fundamental problem. We are what train them, we are what they have to contend with in the real world, and we don't understand what's going on "under the covers", with them or with us.

FSD still hasn't fully achieved my first law: Never, ever run into things!

 

[65SoYoLO]

This should go without saying, the only question is how long will it take? 18 years

FIFY

Sponsored